Predicting Anti-Asian Hateful Users on Twitter during COVID-19
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Predicting Anti-Asian Hateful Users on Twitter during COVID-19 Jisun An1 , Haewoon Kwak1 , Claire Seungeun Lee2 , Bogang Jun3 , Yong-Yeol Ahn4 1 School of Computing and Information Systems, Singapore Management University 2 School of Criminology and Justice Studies, University of Massachusetts Lowell 3 Department of Economics, Inha University 4 School of Informatics, Computing, and Engineering, Indiana University, Bloomington {jisunan,hkwak}@smu.edu.sg, claire_lee@uml.edu bogang.jun@inha.ac.kr, yyahn@iu.edu Abstract 2020) and negativity towards China on Twitter and Reddit (Schild et al., 2020). Many Asians have We investigate predictors of anti-Asian hate expressed an increased level of anxiety and de- among Twitter users throughout COVID-19. With the rise of xenophobia and polarization pression due to the recent development of racial that has accompanied widespread social me- animus (Wu et al., 2021). As an attempt to investi- dia usage in many nations, online hate has be- gate online hate towards Asians during COVID-19, come a major social issue, attracting many re- several studies have introduced new datasets and searchers. Here, we apply natural language methods for hate detection towards Asians (Ziems processing techniques to characterize social et al., 2020; Vidgen et al., 2020). media users who began to post anti-Asian hate messages during COVID-19. We compare While most studies focus on the content—e.g., two user groups—those who posted anti-Asian whether a given text is a hate speech or not, user- slurs and those who did not—with respect to a level analysis—e.g., whether a given user will rich set of features measured with data prior to post hateful content or not—is surprisingly under- COVID-19 and show that it is possible to pre- explored, although this question can bring useful dict who later publicly posted anti-Asian slurs. Our analysis of predictive features underlines and important insights. First of all, considering the potential impact of news media and infor- the fact that only a handful of users produce the mation sources that report on online hate and vast majority of misinformation (CCDH, 2021) and calls for further investigation into the role of hateful content (§4.2.1), focusing on users may be polarized communication networks and news an efficient way to tackle online hate. For instance, media. identifying ‘influential’ users who tend to produce a large volume of such content and have the capacity 1 Introduction to be the center of the discussion can lead to a better The novel coronavirus pandemic (COVID-19) pro- intervention than identifying individual instances vides a unique opportunity to study the develop- of hate speech. Second, understanding the risk ment of targeted racial animus at an unprecedented factors for hate speech may also provide an oppor- scale. Since the first known case of COVID-19 tunity to nudge the users and keep them from being was reported in Wuhan, China (Mallapaty, 2021), radicalized. This pathway is only possible when Asians have been a target of online and offline we examine individual-level risk factors. Finally, hate. Multiple surveys have shown an increase because there tends to be much more data avail- in anti-Asian attitudes among Americans, which able about individual users than individual tweets has manifested in xenophobic behaviors, that range or posts (Qian et al., 2018), we may be able to de- from not visiting Asian restaurants or changing velop a more accurate model as well as in-depth seats to avoiding Asians in public places (Dhanani insights into the development pathways of online and Franz, 2020; Croucher et al., 2020; Reny hate. At the same time, we emphasize that care and Barreto, 2020) to verbal and physical harass- must be taken when taking user-based approaches ment (CSHE, 2021; AAPI, 2021). because the prediction of future offenses and misbe- Online platforms, social media in particular, haviors can potentially lead to algorithmic bias as have been an exemplification, rather than an excep- shown in the case of recidivism prediction (Angwin tion, of anti-Asian hate, hosting plenty of hateful et al., 2016). Thus, our work should be considered content. Recent work has reported a striking in- an early step towards understanding online hate, crease of anti-Asian slurs on Twitter (Ziems et al., and we caution that the translation of our results 4655 Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4655–4666 November 7–11, 2021. ©2021 Association for Computational Linguistics
into practice would require much more nuanced et al., 2017; Waseem and Hovy, 2016; Waseem, investigation and decision-making. 2016; Founta et al., 2018), YouTube (Salminen In this paper, we tackle online hate towards et al., 2018), online games (Kwak et al., 2015), Asians by focusing on users, particularly focus- and many other online platforms (Zannettou et al., ing on the risk indicators of hate towards Asians. 2018), different types of online hate have been in- We examine users’ language use in their tweets and vestigated. While leveraging language to detect information sources that they follow on social me- hate speech is not new (Schmidt and Wiegand, dia. As polarization and echo chamber have been 2017), user-level modeling for detecting a future commonly observed on social media (Garimella (event-driven) risk of posting hate speeches is rela- and Weber, 2017; An et al., 2019), it is reasonable tively unexplored. Among a few, Almerekhi et al. to expect that users would choose to whom they (2020) identify linguistic markers that trigger toxi- listen, often in the way to strengthen their biases. city in online discussions, Ribeiro et al. (2018) de- Thus we expect to see signals from the identities of tect the users who express online hate by checking the followings of users. Although our retrospective their profiles or linguistic characteristics, Qian et al. case-control design does not allow strong causal (2018) incorporate users’ historical tweets and user inference (i.e., we cannot discern whether a user is similarities for hate speech detection, and Lyu et al. affected by whom they follow or their following is a (2020) compares hate instigators, targets, and gen- manifestation of their existing bias), our study may eral Twitter users by self-presentation, Twitter visi- shed light on potential mechanisms and pathways bility, and personality trait. Lyu et al. (2020) char- towards online hate. Also, while our work focuses acterize Twitter users who use controversial terms on online hate, it would provide helpful insights for associated with COVID-19 (e.g., “Chinese Virus”) modeling offline hate crimes as well (Relia et al., and those who use non-controversial terms (e.g., 2019). “Corona Virus”) in terms of demographics, pro- Our contributions can be summarized as two- file features, political following, and geo-locations. fold. First, we analyze user-level predictors of hate Unlike previous studies, our work focuses on not speech towards Asians. We study the impact of only user features but also on the content posted both linguistic and information sources, namely (i) before they start to use Asian slurs, studying the Content features (social media posts published be- development of anti-Asian hate attitudes. fore COVID-19) and (ii) Content-agnostic features Hate towards Asians during COVID-19 has ap- (what kind of information a user is exposed to, how peared through diverse forms. For example, anti- a user interacts with the platform and other users, Asian slurs are increasingly used (Schild et al., etc). We further study the level of expressed hate. 2020). Also, over 40% of survey respondents in Second, we will release our dataset (features) and the U.S. would engage in discriminatory behavior the code necessary to replicate our results1 . towards Asians due to fear of COVID-19, lack of knowledge about the virus and trust in science, and 2 Related Work more trust in Donald Trump (Dhanani and Franz, Racial animus has been widely studied from the per- 2020). Reny and Barreto (2020) reported that xeno- spective of its impact on individuals’ health (Sellers phobic behavior as well as concerns about the virus et al., 2003), economic development (Card et al., is associated with anti-Asian attitudes. 2008), voting (Stephens-Davidowitz, 2014) and The context and role of social media in online social unrest (BBC, 2020b). Understanding how and offline hate have been argued to be impor- racial animus is developed is essential to prevent tant. Croucher et al. (2020) examined a link be- the risk of further intensifying racial animus and tween social media use and xenophobia toward reduce its harm to society. Asians. Ziems et al. (2020) demonstrated that they Online hater has been studied actively in recent could identify hate and counterhate tweets with years. On internet forum (de Gibert et al., 2018), an AUROC of 0.85. Vidgen et al. (2020) built an Wikipedia (Wulczyn et al., 2017), News media annotated corpus for four Asian prejudice-related comments (Coe et al., 2014), Twitter (Davidson classes, including criticism without being abusive. 1 Their best model achieved a macro-F1 score of The data and the corresponding code are avail- able at https://github.com/anjisun221/ 0.83, but the only 59.3% of hostility tweets were Anti-Asian-Slurs correctly identified. 4656
3 Data Ref Hate Low H High H Total We first identify those who express hate towards Users 2,443 1,899 1,316 583 4,342 Tweets 250k 8.2M 4.8M 3.4M 10.1M Asians by collecting data from Twitter in three steps: 1) compile a list of anti-Asian slurs; 2) col- Table 1: Dataset overview lect tweets with any of the anti-Asian slurs; and 3) identify those who have used anti-Asian slurs after the pandemic began (we call them hateful COVID-19 case as the start date of COVID-19. users) and collect their historical tweets from June Refining reference users. As explained above, 8, 2019 to May 8, 2020. Then, as a control set, our reference users are randomly selected from a we randomly sample users who have tweets about large collection of COVID-19-related tweets. We COVID-19 (we call them reference users) and col- find that 19 reference users (0.6%) have used anti- lect their historical tweets during the same period. Asian slurs before COVID-19. We exclude those Anti-Asian slurs. We compile a list of anti- users for the rest of the study. Asian slurs by combining 1) Wikipedia’s list of ethnic slur words (Wikipedia, 2021) that includes Refining hateful users. Some hateful users have ‘chink,’ ‘chinazi,’ and ‘chicom,’ and 2) a set of used anti-Asian slurs before COVID-19. Through COVID-19 specific anti-Asian slurs, such as ‘wu- manual examination, we notice that most of them flu’ and ‘kungflu’ (Schild et al., 2020). The full list are activists in Hong Kong, expressing negativity of 33 anti-Asian slurs is in Appendix A. through slur words targeting China. We exclude these users from further analyses. Hateful users. Using Twint (Poldi, 2021), in By contrast, the use of slur words like “wuflu” May 2020, we collect tweets that contain any of and “kung flu” began to increase starting from the anti-Asian slurs from December 31, 2019 to March 9, 2020, a day when Italy extends emer- May 1, 2020, resulting in 190,927 tweets posted by gency measures nationwide (BBC, 2020a), and 120,690 users. We then consider those who 1) live showed a sharp spike on March 16, 2020, when in the U.S. and 2) posted at least two tweets with the former US president Donald Trump referred to anti-Asian slurs. We use self-declared locations in COVID-19 as “Chinese Virus” on Twitter (News, user profiles to infer their state-level location and 2020). Since then, all anti-Asian slurs becomes exclude users without identified locations. This widely used. We focus on these users who turned leaves us with 3,119 hateful users. hateful after the COVID-19. Reference users. As a control set, we construct Low- vs High-level Hateful users. We further a set of non-hate users. Using Twitter’s Streaming divide hateful users into two groups based on the API, we collect 250M tweets that include COVID- level of expressed haterism towards Asians. In our 19 related keywords (e.g., “covid” and “coron- data collection, there are 8,769 tweets that contain avirus”) from January 13 to April 12, 2020. The at least one slur word. The average number of full list of COVID-19 keywords used for this data tweets with a slur word per user is four (min: 1, collection can be found in Appendix B. From this median: 3, max: 126). Thus, we further divide dataset, we randomly select 3,119 reference users, users into two groups based on the average tweets whose location can be detected at state-level and with slurs: Low-level hateful users with less than who have not used any anti-Asian slurs. or equal to three tweets with anti-Asian slurs and Historical tweet collections. For the two user High-level hateful users with at least four tweets groups, we collect their historical tweets posted for with anti-Asian slurs. 11 months from June 8, 2019 to May 8, 2020 using Bot detection. We further remove ‘bot’ accounts Twint package (Poldi, 2021). In total, we collect to better capture genuine human behaviors. We 18,952,895 tweets where 15.91M (83.94%) tweets use Botometer API (Davis et al., 2016), a popu- of hateful users and 3.04M (16.06%) tweets of lar supervised machine learning tool that checks reference users. We find 15,728 tweets (0.00083%) Twitter accounts for possibly automated activity by containing anti-Asian slurs. using features including content, network structure, Pre- and Post-COVID-19 tweets We use De- and profile. Given a Twitter account, Botometer cember 31, 2019 when China confirmed the first API returns a score from 0 to 1, where 1 is the most 4657
likely to be a bot. We use the fifty-percent threshold from Twitter Decahose data. The unigrams are then (Botometer score = 0.5), which has proven effec- ranked by their estimated z-scores. tive in prior studies (Davis et al., 2016), to remove potential bot accounts. Data summary. As a result, our data collec- tion for further analyses contains 8,201,510 and 2,498,246 tweets posted by 1,899 hateful users and 2,443 reference users, respectively. We also collect their network information—whom they follow by (a) (Pre) Hateful users (b) (Pre) Reference users using Twitter REST API. Details are in Table 1. 4 Comparative Exploration of Hateful and Reference Users Prior to predicting future hate behaviors of users, we first conduct a comparative exploration between (c) (Post) Hateful users (d) (Post) Reference users hateful users and reference users to better under- stand their discriminating characteristics. Figure 1: Over-represented unigrams of Hateful users and Reference users for Pre- and Post-COVID-19 4.1 How much and what they write 4.1.1 Twitter Activity Figure 1 shows the 100 most over-represented The first noticeable difference between hateful unigrams from each corpus. For hateful users be- users and reference users is their activity level fore COVID-19 (Figure 1(a)), US politics-related in terms of the number of written tweets after words (e.g., obama, biden, clinton, conservative, COVID-19. On average, both user groups in- and democrat) and international political issue- creased their activities, and the increase is more related words (e.g., iran, border, war, and ukraine) considerable for hateful users. We evaluate the are presented. Some words indicate that these users statistical significance by 1,000 bootstrapped sam- are likely to be right-wing: 1) maga (“Make Amer- ples. The bootstrapped average percent increase ica Great Again”) is a campaign slogan used in among hateful users is much higher than reference American politics popularized by Donald Trump, users: 90.12 % (±0.419 %, 95% CI) vs. 24.52 % 2) terms like leftists and socialist, and 3) right-wing (±0.107 %, 95% CI). Furthermore, we observe that news media, including breitbartnews, gatewaypun- the increase in activity after COVID-19 is driven dit, and dailycaller. Another interesting word set more by those High-level hateful users than Low- contains names: soros (likely George Soros), omar level ones. The bootstrapped average percent in- (likely Ilhan Omar), schiff (likely Adam Schiff), crease among High- and Low-level hateful users and nancy (likely Nancy Pelosi), who opposed are 118.97 % (±0.977 %, 95% CI) and 77.22 % Donald Trump. Also, a hashtag ‘#votebluenomat- (±0.381 %, 95% CI), respectively. terwho’ seems to be used by hateful users. The reference users’ prominent words before 4.1.2 Representative Words of User Groups COVID-19 (Figure 1(b)) look more casual than As an exploratory analysis, we identify the repre- hateful users, which also show the validity of sentative words of user groups by using the log- our reference sample to some extent. The over- odds ratio proposed in (Monroe et al., 2008). For represented words are related to sports and enter- each group, we aggregate all their pre-COVID-19 tainment (e.g., nba, football, game, and music), and post-COVID-19 tweets separately, creating the and K-pop (e.g., bts, nct, and exo), which is one of four corpora. To remove rare jargon, we elimi- the globally popular topics on Twitter (Kim, 2021). nate terms that appear less than 10 times. We then Also, words with positive connotations (e.g., love, extract all unigrams and compute their log-odds happy, thank, congrats, and excited, lmao (“laugh- ratio. As the prior, we compute background word ing my ass off”)) are appeared. We find some N- frequency on two separate random Twitter datasets word expressions from this group, suggesting the for pre-COVID-19 and post-COVID-19, sampled presence of more black Twitter users in this group. 4658
After COVID-19, hateful users seem to actively To examine what information users are exposed to, write tweets about COVID-19 (Figure 1(c)), includ- we opt for analyzing news URLs shared by users. ing China, Chinese, virus, flu, death, and Wuhan. In doing so, we use media-level factuality ratings Words related to infodemic, such as propaganda, on a 7-point scale (Questionable-Source, Very-Low, fake, wrong, lying, and truth, are also presented, Low, Mixed, Mostly-Factual, High, and Very-High) which is well aligned with a previous report (Cinelli and bias ratings on a 7-point scale (Extreme-Left, et al., 2020). Left, Center-Left, Center, Center-Right, Right, and Lastly, Figure 1(d) shows reference users’ over- Extreme-Right) annotated by the Media Bias/Fact represented words after COVID-19. We do not see Check (MBFC) (Zandt, 2015). many changes for this group but note two particular words: kobe and medicareforall. Kobe (Bryant) is # of shared URLs (per user) Hateful # of shared URLs (per user) 6 10.0 Hateful Reference Reference an NBA superstar who passed away in January 4 7.5 5.0 2020. Since ‘medicareforall’ is one of the key 2 2.5 slogans of the Democrats, this suggests that more 0 0.0 eft ft ter r ht ht r left-wing users may be present in the reference h w h Mi l d w le nte nte tua Le Hig Hig xe Rig Rig Lo Lo ab L en Ce Ce ac me n ry ft C ry tio me yF Ve Ve ht tre es Le tre Rig stl Ex Qu Mo Ex users. Media Bias Media Factuality (a) Bias (b) Factuality 4.1.3 How their tweets are engaged by others We examine how actively other users engage in # of shared URLs (per user) # of shared URLs (per user) High-level Hateful 25 High-level Hateful Low-level Hateful Low-level Hateful 15 20 hateful users’ tweets. The average retweet (like) 10 15 counts of hateful users is 1.88 (7.88), while that 5 10 5 of reference users is 1.47 (7.68). We run a Mann- 0 0 h w h Mi l d w le Whitney’s U test to evaluate the difference and eft ft ter r ht t r tua igh Hig nte nte Hig xe Lo Lo Le ab Rig eL en ac R Ce on Ce ry ry ft C m yF me Ve ti Ve ht tre es Le stl tre Rig Qu Ex Mo Ex find a significant effect of group for retweet counts Media Bias Media Factuality (U = 6.05, p < 0.001), but not for like counts. (c) Bias, High vs Low (d) Factuality, High vs Low Furthermore, high-level hateful users tend to have higher retweet counts but lower like counts than Figure 2: Shared News Media before COVID-19 low-level users. The average retweet (like) counts are 2.97 (1.01) and 1.11 (6.29) for high-level and We compare the average number of shared URLs low-level hateful users, respectively. The differ- for each of categories in media bias (Figure 2(a)) ences of retweet and like counts are statistically and factuality (Figure 2(b)) between hateful and significant (p < 0.001 and p < 0.05, respectively). reference users from their pre-COVID-19 tweets. Lastly, tweets with anti-Asian slurs tend to be Across all categories, hateful users share more more retweeted and liked than other tweets of hate- news URLs than reference users. While the rep- ful users. The average retweet (like) counts are 4.52 resentative word analysis hints that hateful users (10.87) for tweets with anti-Asian slurs and 1.87 are more likely to be right-wing, their shared news (7.88) for other tweets. We find that the like count are fairly diversified. While hateful users share difference is statistically significant (p < 0.001) URLs from credible news media, they also share but not the retweet count. The results indicate that many news URLs from less credible news media. the tweets posted by hateful users are more likely When comparing high-level and low-level hateful to be propagated and thus get exposed to the target users, active news sharing behavior of hateful users group. Moreover, it suggests that there may exist mostly comes from high-level hateful users. positive feedback for hateful tweets—expressing hate tends to increase engagement, which in turn 4.2.2 Followings may nudge users to post more hate tweets. The accounts following is yet another proxy of in- formation sources. We examine Twitter accounts 4.2 What They Consume and Share that are the most followed by each group. For refer- 4.2.1 Shared News Media ence users, the top 5 are ‘BarackObama’, ‘realDon- News is known to be powerful in shaping people’s aldTrump’, ‘nytimes’, ‘AOC’, and ‘cnnbrk’, while opinions and behaviors (McCombs and Valenzuela, those by hateful users are ‘realDonaldTrump’, ‘Re- 2020). A potential bias or factuality of news report- alJamesWoods’, ‘POTUS’, ‘DonaldJTrumpJr’, and ing thus can influence one’s attitude towards Asians. ‘TuckerCarlson’. In the top 50 Twitter accounts 4659
followed by the two groups, only 5 are in common, readability), bias (Recasens et al., 2013; Mukherjee indicating that the information sources of the two and Weikum, 2015), and morality (Graham et al., groups are distinctive. Among hateful users, the 2009; Lin et al., 2018). These features are known to overlap between low- and high-level hateful user be indicative for detecting fake information and the groups is high—45 of the top 50 are shared. political bias of news sources. We expect that they help to capture aspects of factuality or political 4.3 Summary bias of hateful users’ tweets. For each user, we In sum, hateful users, as a group, exhibit noticeable obtain her NELA features by averaging the NELA differences in comparison with the reference users, features of her tweets. including 1) being more active on Twitter during Psycholinguistics Features: We extract features COVID-19, 2) using more words about politics (es- that relate to emotional and psychological charac- pecially right-wing), 3) sharing significantly more teristics of users by using Linguistic Inquiry and URLs of news media, and 4) having distinctive Word Count (LIWC) (Pennebaker et al., 2015). information sources. High-level hateful users can be further differen- Embedding Features: We encode each tweet tiated from low-level ones by the following fea- using Sentence BERT (SBERT) (Reimers and tures: 1) they increased Twitter activity even more Gurevych, 2019) for the following reasons: 1) the (about 54% more than low-level hateful users) after distribution of the number of tweets per user is COVID-19, 2) their tweets tend to be more shared highly skewed, and 2) tweet has a sentence-like and liked, and 3) they share more URLs published structure and length. Thus, we opt for SBERT by news media of extreme bias and low factuality. and average the SBERT representations across the sample tweets to obtain a user-level representation. 5 Predicting Hateful Users The averaged user-level representation is a 768- Our comparative exploratory analysis indicates that dimensional vector. there exist significant differences between these Shared News Media: We also consider news groups, raising an important question: can we pre- media shared by users as features because they dict those who turned hateful by only using the can be a proxy for information sources that users information available prior to the pandemic? The consume and propagate. Using seven categories possibility of such prediction, combined with the of bias and factuality of news media introduced in analysis of prominent predictors, may help us un- §4.2.1. derstand the pathways into xenophobic haterism and design potential interventions. 5.1.2 Content-agnostic features We model users based on how they interact with 5.1 Features others and how they portray themselves to their For each user, we extract 1) content features and 2) audiences by using their Twitter profiles. content-agnostic features for training a classifier to Twitter Statistics: We use basic information ex- predict the future hate expression towards Asians. tracted from a Twitter profile, such as 1) whether a 5.1.1 Content Features user’s account is verified by Twitter; 2) the number of days since the account was created; and 3) four From the over-represented words in Figure 1, we Twitter statistics including the number of followers, observe that word usage is distinct between the two followings, tweets, and favorites. user groups. We thus extract content features from the pre-COVID-19 tweets of every user. Since the Twitter Profile Description: Since most of the distribution of the number of tweets per user is Twitter profile description (bio) is short and has skewed, we sample 198 tweets per user, a median a sentence-like structure, we encoded the user’s number of pre-COVID-19 tweets per user, and ex- profile description using SBERT, like Baly et al. tract content features from the sampled 198 tweets. (2020) did for obtaining followers’ representations. NELA Features: We extract following linguistic Twitter Following: We consider those who fol- features by using the News Landscape (NELA) lowed by a given user as features, which capture toolkit (Horne et al., 2018): text structure (e.g., whom they listen to (information source). First, POS tags), sentiment, subjectivity, complexity (e.g., we identify the top 50 Twitter accounts followed 4660
by each user group and obtain their union, which features generally shows improvements, except for results in 95 Twitter accounts. Then, for each user, one case—profile description features yield loss in we check whether one follows each of them and performance when added to the following features create a 95-dimensional vector as the following (row 7). Twitter statistics and profile description features. improve the performance when combined with the following features, yielding the best performance 5.2 Experimental Setup among content-agnostic features (row 8). We evaluate (i) content and (ii) content-agnostic Rows 9-12 show that average embeddings of features separately and in combinations. We train tweets by SBERT (row 12) work better than NELA XGBoost classifier for predicting whether or not a features (row 10) or LIWC features (row 11). We user will express hate towards Asians in the future. note that a combination of shared news media, Although we could have adopted other models to NELA, and LIWC features shows improvements improve the performance, our focus is on under- (row 19), but they are worse than using SBERT fea- standing risk factors rather than building the best tures. While news media features alone do not yield prediction models. We thus chose XGBoost, which good performance, it gives a sizable improvement is highly robust across different data and problems, by +0.55 macro-F1 points (row 15) when added to regularly outperforms more sophisticated models. the SBERT features, which is the best performance We perform an incremental ablation study by com- among content features. bining the best features from (i) and (ii) to achieve Finally, rows 24 and 26 show the results when even better results. We split data in 80:20 ratio for combining content and content-agnostic features. training and testing. With training data, we train The best result is achieved using all features (row and evaluate an XGBoost model using different 26). This combination improves over using in- features and feature combinations. At each itera- formation from contents only (row 15) by +2.47 tion of the 5-fold cross-validation, we perform a macro-F1 points. The result indicates that not grid search to tune the hyper-parameters of our XG- only users’ tweets but their information sources Boost model, which are the maximum tree depth have strong predictive power to identify those who and the minimum child weight that controls for would express hate against Asians after COVID- complexity and conservativeness, respectively. We 19, demonstrating the advantage of the user-level use the learning rate of 0.1, gamma of 0.1, and col approach than tweet-level approach. tree of 0.8. In the search process, we optimize for macro-average F1 score, i.e., averaging over the 5.4 T2: High-level Hateful User Prediction classes, since our dataset is not balanced for both tasks. Finally, we evaluate the model on the unseen The second task is to predict whether a user would testing data. We report both macro F1 score and turn into high-level hateful users against Asians. Ta- accuracy and compare our result with the majority ble 2 (Column 7 & 8) shows the evaluation results. class baseline. We note that the dataset for this task is imbalanced (See Table 1), yielding high accuracy (71.05) for 5.3 T1: Hateful User Prediction our baseline, majority class model. Overall, the Table 2 shows the evaluation results for future hate performance of this task is not as high as hateful prediction grouped by feature categories. For each user prediction, reflecting the difficulty of this task. category, the upper rows correspond to an individ- Similar to the result of hateful user prediction, ual set of features, while the lower ones show their rows 2-4 suggest that ‘whom they follow’ is more combinations. important than Twitter statistics or profile descrip- Rows 2-4 show that whom they follow (row 4) is tion. Rows 5-9 suggest that a combination of Twit- the most useful feature among the content-agnostic ter statistics and following features (row 6) result features. As followings determine what kinds of in the highest performance among content-agnostic information a user would be exposed to, this result features. Rows 9-12 indicate that LIWC features indicates two groups are likely to be exposed to a (row 11), which capture the psycholinguistic char- very different set of information at least on Twitter. acteristic of users, are better than embedding and Rows 5-8 show the results of the models that NELA features. Combining shared news media, combine Twitter statistics, profile description, and NELA, and LIWC (row 19) shows a slight im- following features. Combining content-agnostic provement over the model using LIWC features. 4661
Hateful User High-level Hateful User Model # Features Dim. Macro F1 Accuracy Macro F1 Accuracy Baselines 1 Majority class 36.15 56.62 41.54 71.05 A. Content 2 Twitter Statistics 6 67.87 68.58 46.88 67.37 -agnostic 3 Profile Description: SBERT 768 64.60 66.05 46.88 67.37 features 4 Following 95 80.04 81.01 51.41 63.95 5 Twitter Stat. + Prof. 774 72.52 73.30 49.29 68.68 6 Twitter Stat. + Fol. 101 80.63 81.13 58.80 69.21 7 Twitter Prof. + Fol. 863 79.22 80.21 50.29 67.11 8 Twitter Stat. + Prof. + Fol. 869 80.98 81.70 51.36 67.37 B. Content 9 Shared News media 14 67.87 71.35 49.82 68.68 features 10 Tweets: NELA 85 74.20 74.68 54.44 66.32 11 Tweets: LIWC 73 76.36 76.75 55.06 67.63 12 Tweets: SBERT 768 81.55 81.93 49.89 64.47 13 Media + NELA 99 75.38 76.06 55.98 68.42 14 Media + LIWC 87 78.48 78.83 56.32 68.42 15 Media + SBERT 782 82.10 82.51 48.47 64.21 16 NELA + LIWC 158 77.64 78.02 53.60 66.58 17 NELA + SBERT 853 81.68 82.05 53.23 66.58 18 LIWC + SBERT 841 81.43 81.82 47.43 63.95 19 Media + NELA + LIWC 172 78.62 79.17 55.98 68.42 20 Media + NELA + SBERT 867 81.95 82.39 50.12 66.84 21 Media + LIWC + SBERT 855 81.17 81.59 51.33 66.05 22 NELA + LIWC + SBERT 926 81.24 81.70 53.78 67.89 23 Tweets: ALL 940 81.49 81.93 47.08 62.63 Combinations 24 (Task 1) A+B: rows 8 & 15 1651 84.03 84.58 - - 25 (Task 2) A+B: rows 6 & 19 273 - - 54.58 66.05 26 All features 1833 84.57 85.04 49.17 66.05 Table 2: Ablation study of the proposed features for hateful user prediction (Task 1) and high-level hateful user prediction (Task 2). The Dim. column indicates the number of features (dimensions) used for each experiment. Comparing the best model of content-agnostic gust, gross) (Curtis, 2011). Examining the LIWC features (row 6) with that of content features (row based model (row 11), hateful users tend to use 19), unlike the results for hateful users predic- more ‘they’ than ‘I’ or ‘we’ and use more words tion, content features perform worse than content- relating to power, risk, religion, male, wear, non- agnostic features. In other words, it is hard to distin- fluencies (uh, rr*) and less leisure and work related guish high-level hateful users from low-level ones words. Lastly, linguistic features that predict high- based on what they write (content). Instead, what level hateful users are: using more negative words information they subscribe to (following) has more and ‘I’ and less all capitalized words, punctuation, explanatory power for the level of hate expression. positive and anxiety words, and internet slang. 5.5 Important Linguistic Features 6 Discussion and Conclusion We examine the important linguistic features us- We presented a study on predicting users who ing SHAP (Lundberg et al., 2020). Examining would express hate towards Asians during COVID- the top 20 most important features of the NELA- 19 by using their language use and information based model (row 10), hateful users tend to use sources before COVID-19. We modeled a user more strong negative words and ‘there’ while us- by a rich set of features derived from 1) contents ing less punctuation, positive words, and plural published by the user and 2) content-agnostic di- nouns. Two moral dimensions, Care/Harm and mensions, including their Twitter statistics, profile Purity/Degradation, are helpful to identify hateful description, and followings. For hateful user predic- users. Hateful users use more words relating to tion task, our evaluation results showed that most ‘Harm’ (e.g., war, kill) and ‘Degradation’ (e.g., dis- features have a notable impact on prediction per- 4662
formance, which are tweets represented by SBERT, and the set of content with anti-Asian slurs would followings, LIWC, NELA, shared news media, not completely overlap, we argue that because of Twitter statistics, and profile description (in this the aforementioned reasons, our target population order). For high-level hateful user prediction, fol- (those who used slurs instead of those who posted lowing features turn out to be more important than “hateful content”) is still a useful operationalization content features. Moreover, embedding features of anti-Asian hate. Furthermore, this operational- are worse than NELA or LIWC features, indicating ization allows a highly transparent and unambigu- that linguistic styles and information sources are ous definition of the population. If we target the crucial for predicting levels of hate towards Asians. population with “hate towards Asians,” the oper- ationalization would inevitably require adopting Our retrospective case-control design enabled us methods that are much more difficult to understand to study the distinctive features of hateful users in and evaluate. While the keyword-based method comparison with reference users. We reveal their can result in a skewed dataset due to the oversam- individual importance and contribution, providing pling of certain keywords (e.g., n-word) (Vidgen interpretable insights. In contrast to previous work and Derczynski, 2020), since we collect all tweets focusing on user features on hate content, our study containing Asian slurs to study users, not tweets, sheds light on potential mechanisms and pathways we believe the sampling bias would not be a critical (risk factors) towards online hate. In particular, our issue in our study. Furthermore, we would like to finding that one feature, following, has a strong note that we exclude users who have used Asian predictive power provides compelling sociologi- slurs before COVID-19 to ensure that our data cap- cal relevance. This finding hints at social factors tures users who newly developed anti-Asian atti- (which communities they belong to) potentially be- tudes. ing dominating factor in the development of racial For future work, we plan to address the predic- hatred, suggesting a strong link between social po- tion task by ordinal regression that can inherently larization and xenophobia and calling for actions model the level of hate. We are also interested of social media companies and our society. in characterizing hate towards Asians in other lan- There are some limitations to our work. First, guages. Finally, we want to go beyond an observa- this work is an observational study. Since we model tional study and attempt to find a potential causal information sources based on shared URLs and fol- relationship between information sources (biased, lowings, we could not know information sources less credible information) and online hate. that are neither shared nor followed by users. Sec- ond, as our study is based on the US and anti-Asian Acknowledgements behavior, further studies would require to gener- This research was supported by the Singapore alize our findings for other countries or for other Ministry of Education (MOE) Academic Research ethnic minorities. Third, our target population is Fund (AcRF) Tier 1 grant, the DS Lee Foundation those who used anti-Asian slurs, and thus, other Fellowship awarded by the Singapore Management forms of anti-Asian hate may not be included in University, and the University of Massachusetts this work. However, we argue that not only that Lowell’s COVID-19 Seed Award (Mapping Cy- the keyword-based method is a widely adopted berhate toward Asian Americans in Digital Media approach for studying anti-Asian attitudes during Platforms in the COVID-19 Era, 2020-2022). The COVID-19 (Lu and Sheng, 2020; Schild et al., funders had no role in study design, data collection 2020; Lyu et al., 2020), but also the population and analysis, decision to publish, or preparation of using Asian slurs itself is of great importance be- the manuscript. cause anti-Asian slurs 1) are unambiguously pe- jorative (Camp, 2013) and context-independent Ethical Considerations (Hedger2010); 2) had not been commonly used un- like other racial slurs before the pandemic (Schild While this work provides important insights on the et al., 2020); and 3) have not been reclaimed by the development pathways of online hate, it also brings Asian American community (Croom, 2018); and ethical implications by potentially being able to 4) their prevalence has been linked to offline be- reveal who may turn to haters. We note that all haviors and hate crimes during COVID-19 (Lu and data collected in this study is publicly available, Sheng, 2020). Although the set of hateful content and we did not attempt to use any individual-level 4663
demographic information in our study. Yet, we BBC. 2020a. Coronavirus: Italy extends emergency are able to predict whether an individual turns into measures nationwide. hateful after the COVID-19 at F1 = 85.57% with BBC. 2020b. George Floyd: ’pandemic of racism’ led simple machine learning methods based on pub- to his death, memorial told. licly available Twitter information. The risk of such user-based predictive tools for future offenses Elisabeth Camp. 2013. Slurring perspectives. Analytic and misbehaviors is often underestimated and can Philosophy, 54(3):330–349. potentially lead to algorithmic bias, as shown in David Card, Alexandre Mas, and Jesse Rothstein. 2008. the case of recidivism prediction (Angwin et al., Tipping and the dynamics of segregation. The Quar- 2016). Thus, we once again emphasize that our terly Journal of Economics, 123(1):177–218. results should be considered as a first step for fight- CCDH. 2021. The disinformation dozen: Why ing online hate and further studies are required for platforms must act on twelve leading online anti- translating it for practical use. This work is exempt vaxxers. from the requirement for IRB review and approval (Reference #11649). Matteo Cinelli, Walter Quattrociocchi, Alessandro Galeazzi, Carlo Michele Valensise, Emanuele Brug- While sharing tweet IDs is a common practice noli, Ana Lucia Schmidt, Paola Zola, Fabiana Zollo, of studies using Twitter data, there is a risk to share and Antonio Scala. 2020. The COVID-19 social me- tweet IDs in this work due to the sensitivity of dia infodemic. Scientific Reports, 10(1):1–10. the dataset. For example, even if a user in the Kevin Coe, Kate Kenski, and Stephen A. Rains. 2014. “hateful” user set deleted all the hate tweets, others Online and Uncivil? Patterns and Determinants of still may be able to see whether a particular user Incivility in Newspaper Website Comments. Jour- posted anti-Asian slurs or not. This scenario may nal of Communication, 64(4):658–679. become possible if someone attempts to combine Adam M Croom. 2018. Asian slurs and stereotypes multiple tweet collections regarding COVID-19. in the USA: A context-sensitive account of dero- Hence, we opt out of sharing tweet IDs. Instead, gation and appropriation. Pragmatics and Society, we share user-level features without revealing any 9(4):495–517. personal information such as username, a profile Stephen M Croucher, Thao Nguyen, and Diyako Rah- description, etc. mani. 2020. Prejudice toward Asian Americans in the COVID-19 pandemic: The effects of social me- dia use in the united states. Frontiers in Communi- References cation, 5:39. AAPI. 2021. Stop aapi hate national report. CSHE. 2021. Fact sheet: Anti-Asian prejudice March Hind Almerekhi, Haewoon Kwak, Joni Salminen, and 2021 by center for the study of hate & extremism. Bernard J Jansen. 2020. Are these comments trigger- ing? predicting triggers of toxicity in online discus- Valerie Curtis. 2011. Why disgust matters. Philosoph- sions. In Proceedings of The Web Conference 2020, ical Transactions of the Royal Society B: Biological pages 3033–3040. Sciences, 366(1583):3478–3490. Jisun An, Haewoon Kwak, Oliver Posegga, and An- Thomas Davidson, Dana Warmsley, Michael Macy, dreas Jungherr. 2019. Political discussions in homo- and Ingmar Weber. 2017. Automated hate speech geneous and cross-cutting communication spaces. detection and the problem of offensive language. In In Proceedings of the International AAAI Confer- Proceedings of the International AAAI Conference ence on Web and Social Media (ICWSM), volume 13, on Web and Social Media, volume 11. pages 68–79. Clayton Allen Davis, Onur Varol, Emilio Ferrara, Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Alessandro Flammini, and Filippo Menczer. 2016. Kirchner. 2016. Machine bias. Botornot: A system to evaluate social bots. In Pro- ceedings of the 25th international conference com- Ramy Baly, Georgi Karadzhov, Jisun An, Haewoon panion on world wide web, pages 273–274. Kwak, Yoan Dinkov, Ahmed Ali, James Glass, and Preslav Nakov. 2020. What was written vs. who Ona de Gibert, Naiara Perez, Aitor Garcia-Pablos, and read it: News media profiling using text analysis and Montse Cuadros. 2018. Hate speech dataset from social media context. In Proceedings of the 58th An- a white supremacy forum. In Proceedings of the nual Meeting of the Association for Computational 2nd Workshop on Abusive Language Online, Brus- Linguistics, pages 3364–3374, Online. Association sels, Belgium. Association for Computational Lin- for Computational Linguistics. guistics. 4664
Lindsay Y Dhanani and Berkeley Franz. 2020. Unex- Smriti Mallapaty. 2021. After the WHO report: what’s pected public health consequences of the COVID-19 next in the search for COVID’s origins. Nature, pandemic: a national survey examining anti-Asian 592:337–338. attitudes in the USA. International journal of pub- lic health, 65(6):747–754. Maxwell McCombs and Sebastián Valenzuela. 2020. Setting the agenda: Mass media and public opinion Antigoni Maria Founta, Constantinos Djouvas, De- (3rd ed). John Wiley & Sons. spoina Chatzakou, Ilias Leontiadis, Jeremy Black- burn, Gianluca Stringhini, Athena Vakali, Michael Burt L Monroe, Michael P Colaresi, and Kevin M Sirivianos, and Nicolas Kourtellis. 2018. Large Quinn. 2008. Fightin’words: Lexical feature selec- scale crowdsourcing and characterization of Twit- tion and evaluation for identifying the content of po- ter abusive behavior. In Twelfth International AAAI litical conflict. Political Analysis, 16(4):372–403. Conference on Web and Social Media (ICWSM). Subhabrata Mukherjee and Gerhard Weikum. 2015. Venkata Rama Kiran Garimella and Ingmar Weber. Leveraging joint interactions for credibility analysis 2017. A long-term analysis of polarization on Twit- in news communities. In Proceedings of the 24th ter. In Proceedings of the International AAAI Con- ACM International on Conference on Information ference on Web and Social Media (ICWSM), vol- and Knowledge Management, pages 353–362. ume 11. NBC News. 2020. Trump tweets about coronavirus us- Jesse Graham, Jonathan Haidt, and Brian A Nosek. ing term ‘Chinese Virus’. 2009. Liberals and conservatives rely on different James W Pennebaker, Ryan L Boyd, Kayla Jordan, and sets of moral foundations. Journal of personality Kate Blackburn. 2015. The development and psy- and social psychology, 96(5):1029. chometric properties of LIWC2015. Technical re- port. Benjamin D Horne, William Dron, Sara Khedr, and Sibel Adali. 2018. Assessing the news landscape: Francesco Poldi. 2021. Twint-Twitter intelligence tool. A multi-module toolkit for evaluating the credibility of news. In Companion Proceedings of the The Web Jing Qian, Mai ElSherief, Elizabeth Belding, and Conference 2018, pages 235–238. William Yang Wang. 2018. Leveraging intra-user and inter-user representation learning for automated YeonJeong Kim. 2021. #kpopTwitter achieves new hate speech detection. In Proceedings of the 2018 record of 6.7 billion Tweets globally in 2020. Conference of the North American Chapter of the Association for Computational Linguistics: Human Haewoon Kwak, Jeremy Blackburn, and Seungyeop Language Technologies, Volume 2 (Short Papers), Han. 2015. Exploring cyberbullying and other toxic pages 118–123, New Orleans, Louisiana. Associa- behavior in team competition online games. In Pro- tion for Computational Linguistics. ceedings of the 33rd annual ACM conference on human factors in computing systems, pages 3739– Marta Recasens, Cristian Danescu-Niculescu-Mizil, 3748. and Dan Jurafsky. 2013. Linguistic models for an- alyzing and detecting biased language. In Proceed- Ying Lin, Joe Hoover, Gwenyth Portillo-Wightman, ings of the 51st Annual Meeting of the Association Christina Park, Morteza Dehghani, and Heng Ji. for Computational Linguistics (Volume 1: Long Pa- 2018. Acquiring background knowledge to im- pers), pages 1650–1659. prove moral value prediction. In 2018 ieee/acm international conference on advances in social net- Nils Reimers and Iryna Gurevych. 2019. Sentence- works analysis and mining (asonam), pages 552– BERT: Sentence embeddings using Siamese BERT- 559. IEEE. networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing Runjing Lu and Yanying Sheng. 2020. From fear and the 9th International Joint Conference on Natu- to hate: How the COVID-19 pandemic sparks ral Language Processing (EMNLP-IJCNLP), pages racial animus in the United States. arXiv preprint 3982–3992, Hong Kong, China. Association for arXiv:2007.01448. Computational Linguistics. Scott M. Lundberg, Gabriel Erion, Hugh Chen, Alex Kunal Relia, Zhengyi Li, Stephanie H. Cook, and Rumi DeGrave, Jordan M. Prutkin, Bala Nair, Ronit Katz, Chunara. 2019. Race, ethnicity and national origin- Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. based discrimination in social media and hate crimes 2020. From local explanations to global understand- across 100 U.S. cities. Proceedings of the Interna- ing with explainable AI for trees. Nature Machine tional AAAI Conference on Web and Social Media, Intelligence, 2(1):2522–5839. 13(01):417–427. Hanjia Lyu, Long Chen, Yu Wang, and Jiebo Luo. 2020. Tyler T Reny and Matt A Barreto. 2020. Xenophobia Sense and sensibility: Characterizing social media in the time of pandemic: othering, anti-Asian atti- users regarding the use of controversial terms for tudes, and COVID-19. Politics, Groups, and Identi- COVID-19. IEEE Transactions on Big Data. ties, pages 1–24. 4665
Manoel Ribeiro, Pedro Calais, Yuri Santos, Virgílio Cary Wu, Yue Qian, and Rima Wilkes. 2021. Anti- Almeida, and Wagner Meira Jr. 2018. Characteriz- Asian discrimination and the Asian-white mental ing and detecting hateful users on Twitter. In Pro- health gap during COVID-19. Ethnic and Racial ceedings of the International AAAI Conference on Studies, 44(5):819–835. Web and Social Media (ICWSM), volume 12. Ellery Wulczyn, Nithum Thain, and Lucas Dixon. 2017. Joni Salminen, Hind Almerekhi, Milica Milenković, Ex machina: Personal attacks seen at scale. In Soon-gyo Jung, Jisun An, Haewoon Kwak, and Proceedings of the 26th international conference on Bernard Jansen. 2018. Anatomy of online hate: De- world wide web, pages 1391–1399. veloping a taxonomy and machine learning models for identifying and classifying hate in online news Dave Van Zandt. 2015. Media Bias/Fact Check. media. Proceedings of the International AAAI Con- Savvas Zannettou, Barry Bradlyn, Emiliano De Cristo- ference on Web and Social Media, 12(1). faro, Haewoon Kwak, Michael Sirivianos, Gianluca Stringini, and Jeremy Blackburn. 2018. What is Leonard Schild, Chen Ling, Jeremy Blackburn, Gian- Gab: A bastion of free speech or an alt-right echo luca Stringhini, Yang Zhang, and Savvas Zannettou. chamber. In Companion Proceedings of the The 2020. “Go eat a bat, Chang!”: An early look on Web Conference 2018, pages 1007–1014. the emergence of sinophobic behavior on web com- munities in the face of covid-19. arXiv preprint Caleb Ziems, Bing He, Sandeep Soni, and Srijan Ku- arXiv:2004.04046. mar. 2020. Racism is a virus: Anti-asian hate and counterhate in social media during the COVID-19 Anna Schmidt and Michael Wiegand. 2017. A survey crisis. arXiv preprint arXiv:2005.12423. on hate speech detection using natural language pro- cessing. In Proceedings of the fifth international workshop on natural language processing for social A Anti-Asian Slur Words media, pages 1–10. We collect tweets that include any of the follow- Robert M Sellers, Cleopatra H Caldwell, Karen H ing slur words: chink, bugland, chankoro, chinazi, Schmeelk-Cone, and Marc A Zimmerman. 2003. gook, insectoid, bugmen, chingchong, chee-chee, Racial identity, racial discrimination, perceived cheechee, cheena, chicom, chinaman, ching cho- stress, and psychological distress among african ing, chingchangchong, ching chang chong, chinki, american young adults. Journal of Health and So- cial behavior, pages 302–317. chinky, chonky, coolie, goo-goos, googoos, gugus, huan-a, jakun, lingling, malaun, panface, wuflu, Seth Stephens-Davidowitz. 2014. The cost of racial an- kung flu, kungflu, yellowman, yellowwoman. imus on a black candidate: Evidence using Google We exclude ‘jap,’ a slur word to express hate search data. Journal of Public Economics, 118:26– 40. against Japanese, as it results in too many false- positive cases in using Twint. Bertie Vidgen and Leon Derczynski. 2020. Direc- tions in abusive language training data, a system- B COVID-19 Related Keywords atic review: Garbage in, garbage out. PloS one, 15(12):e0243300. We collect tweets that match one of the follow- ing keywords: coronavirus, covid, chinese virus, Bertie Vidgen, Scott Hale, Ella Guest, Helen Mar- wuhan, ncov, sars-cov-2, koronavirus, corona, cdc, getts, David Broniatowski, Zeerak Waseem, Austin Botelho, Matthew Hall, and Rebekah Tromble. 2020. N95, epidemic, outbreak, sinophobia, china, pan- Detecting East Asian prejudice on social media. demic, covd. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 162–172, Online. Associa- tion for Computational Linguistics. Zeerak Waseem. 2016. Are you a racist or am I seeing things? annotator influence on hate speech detection on Twitter. In Proceedings of the first workshop on NLP and computational social science, pages 138– 142. Zeerak Waseem and Dirk Hovy. 2016. Hateful sym- bols or hateful people? predictive features for hate speech detection on Twitter. In Proceedings of the NAACL student research workshop, pages 88–93. Wikipedia. 2021. The list of ethinic slurs. 4666
You can also read